Kill forked submission processes after execution

[Sakshamm-Goyal]

Summary

• run each submission in its own session/process group
• kill any descendants left behind after the direct submission process exits
• add a regression that forks a sleeping child, simulates a CPU-limit signal, and verifies the child does not survive

Closes #450

Validation

• fail-first regression: the forked child remained alive before the process-group cleanup
• .venv/bin/python -m pytest -q tests/test_run_program.py tests/test_run_limit.py — 2 passed
• .venv/bin/python -m pytest -q --ignore=tests/test_latex.py --ignore=tests/test_markdown.py --ignore=tests/test_xss.py — 67 passed
• .venv/bin/ruff check .
• .venv/bin/ruff format --check .
• .venv/bin/mypy --non-interactive --config-file mypy.ini -p problemtools

The full local pytest run additionally needs lualatex; without it, the first PDF-render test fails and leaves the process in a removed temporary working directory, which cascades into 10 unrelated render-test failures.

[gkreitz]

Thanks for the PR. This looks quite heavily like a bot, or at least pure LLM code. So I'm guessing PR comments won't get responses, and I'll close this somewhat swiftly if that seems to be the case.

Simon's comments look relevant. Also:

• What other approaches did you consider, and why did you pick this one?
• How does this PR affect pressing ctrl-c to abort a run (please test both with and without -j)?

[Sakshamm-Goyal]

The updated cleanup now:

• uses os.setpgrp() instead of setsid(), so the submission gets a dedicated process group without detaching from the parent session/terminal;
• observes the group leader with wait4(..., WNOWAIT), kills the whole process group, and then reaps the leader (with the macOS wait4 behavior handled explicitly);
• applies the same process-group cleanup when the verifier is interrupted by Ctrl-C;
• treats zombies as still present in the regression check, and writes the child PID through a temporary file plus os.replace() so the test cannot read a partial PID.

I considered killing only the direct PID (leaves forked descendants behind), enumerating descendants manually (racy), and setsid() (heavier and changes session/terminal semantics). A process group gives us the kernel-managed cleanup boundary while keeping the parent session intact.

I also tested Ctrl-C against a real verifyproblem run with both -j 1 and -j 2; in both cases the verifier exited with status 1 and neither the submission nor its forked child remained. The regression keeps the CPU limit real and uses the SIGXCPU path; it only bypasses the unrelated finite-stack-limit setting that macOS rejects.

Validation on the branch:

• pytest -q tests/test_run_program.py tests/test_run_limit.py — 2 passed
• pytest -q --ignore=tests/test_latex.py --ignore=tests/test_markdown.py --ignore=tests/test_xss.py — 67 passed
• Ruff, formatting, compile, and git diff --check all pass

[gkreitz]

Code style is poor, you resolved multiple comments without actually addressing them. Not wasting time on this bot PR, closing.